pipeline_stable_diffusion_3.py

# Copyright 2025 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import inspect
from typing import Any, Callable, Dict, List, Optional, Union

import torch
from transformers import (
    CLIPTextModelWithProjection,
    CLIPTokenizer,
    SiglipImageProcessor,
    SiglipVisionModel,
    T5EncoderModel,
    T5TokenizerFast,
)

from ...callbacks import MultiPipelineCallbacks, PipelineCallback
from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin
from ...models.autoencoders import AutoencoderKL
from ...models.transformers import SD3Transformer2DModel
from ...schedulers import FlowMatchEulerDiscreteScheduler
from ...utils import (
    USE_PEFT_BACKEND,
    is_torch_xla_available,
    logging,
    replace_example_docstring,
    scale_lora_layers,
    unscale_lora_layers,
)
from ...utils.torch_utils import randn_tensor
from ..pipeline_utils import DiffusionPipeline
from .pipeline_output import StableDiffusion3PipelineOutput


if is_torch_xla_available():
    import torch_xla.core.xla_model as xm

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False


#logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers import StableDiffusion3Pipeline

        >>> pipe = StableDiffusion3Pipeline.from_pretrained(
        ...     "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16
        ... )
        >>> pipe.to("cuda")
        >>> prompt = "A cat holding a sign that says hello world"
        >>> image = pipe(prompt).images[0]
        >>> image.save("sd3.png")
        ```
"""

# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
def calculate_shift(
    image_seq_len,
    base_seq_len: int = 256,
    max_seq_len: int = 4096,
    base_shift: float = 0.5,
    max_shift: float = 1.15,
):
    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
    b = base_shift - m * base_seq_len
    mu = image_seq_len * m + b
    return mu


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
    scheduler,
    num_inference_steps: Optional[int] = None,
    device: Optional[Union[str, torch.device]] = None,
    timesteps: Optional[List[int]] = None,
    sigmas: Optional[List[float]] = None,
    **kwargs,
):
    r"""
    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.

    Args:
        scheduler (`SchedulerMixin`):
            The scheduler to get timesteps from.
        num_inference_steps (`int`):
            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
            must be `None`.
        device (`str` or `torch.device`, *optional*):
            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
        timesteps (`List[int]`, *optional*):
            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
            `num_inference_steps` and `sigmas` must be `None`.
        sigmas (`List[float]`, *optional*):
            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
            `num_inference_steps` and `timesteps` must be `None`.

    Returns:
        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
        second element is the number of inference steps.
    """
    if timesteps is not None and sigmas is not None:
        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
    if timesteps is not None:
        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accepts_timesteps:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" timestep schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    elif sigmas is not None:
        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accept_sigmas:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" sigmas schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    else:
        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
        timesteps = scheduler.timesteps
    return timesteps, num_inference_steps


class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin):
    r"""
    Args:
        transformer ([`SD3Transformer2DModel`]):
            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModelWithProjection`]):
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
            specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant,
            with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size`
            as its dimension.
        text_encoder_2 ([`CLIPTextModelWithProjection`]):
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
            specifically the
            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
            variant.
        text_encoder_3 ([`T5EncoderModel`]):
            Frozen text-encoder. Stable Diffusion 3 uses
            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
            [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        tokenizer_2 (`CLIPTokenizer`):
            Second Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        tokenizer_3 (`T5TokenizerFast`):
            Tokenizer of class
            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
        image_encoder (`SiglipVisionModel`, *optional*):
            Pre-trained Vision Model for IP Adapter.
        feature_extractor (`SiglipImageProcessor`, *optional*):
            Image processor for IP Adapter.
        model (`SD3WithRectifiedNoise`, *optional*):
            Optional SD3WithRectifiedNoise model for enhanced noise prediction. If provided, will be used instead of 
            the default transformer for denoising.
    """

    model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae"
    _optional_components = ["image_encoder", "feature_extractor"]
    _callback_tensor_inputs = ["latents", "prompt_embeds", "pooled_prompt_embeds"]

    def __init__(
        self,
        transformer: SD3Transformer2DModel,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModelWithProjection,
        tokenizer: CLIPTokenizer,
        text_encoder_2: CLIPTextModelWithProjection,
        tokenizer_2: CLIPTokenizer,
        text_encoder_3: T5EncoderModel,
        tokenizer_3: T5TokenizerFast,
        image_encoder: SiglipVisionModel = None,
        feature_extractor: SiglipImageProcessor = None,
        model = None,  # 添加 model 参数
    ):
        super().__init__()

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            text_encoder_3=text_encoder_3,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            tokenizer_3=tokenizer_3,
            transformer=transformer,
            scheduler=scheduler,
            image_encoder=image_encoder,
            feature_extractor=feature_extractor,
            model=model,  # 添加 model 参数到 register_modules
        )
        #print(f"VAE is None: {getattr(self, 'vae', None) is None}")
        #if getattr(self, 'vae', None) is not None:
        #    print(f"VAE config block_out_channels: {self.vae.config.block_out_channels}")
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
        #print(f"VAE scale factor: {self.vae_scale_factor}")
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
        self.tokenizer_max_length = (
            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
        )
        self.default_sample_size = (
            self.transformer.config.sample_size
            if hasattr(self, "transformer") and self.transformer is not None
            else 64
        )
        self.patch_size = (
            self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2
        )
        # 添加对 SD3WithRectifiedNoise 模型的支持
        self.model = model

    def _get_t5_prompt_embeds(
        self,
        prompt: Union[str, List[str]] = None,
        num_images_per_prompt: int = 1,
        max_sequence_length: int = 256,
        device: Optional[torch.device] = None,
        dtype: Optional[torch.dtype] = None,
    ):
        device = device or self._execution_device
        dtype = dtype or self.text_encoder_3.dtype
        #max_sequence_length=77
        prompt = [prompt] if isinstance(prompt, str) else prompt
        batch_size = len(prompt)
        
        # print(f"T5处理 - 输入提示: {prompt}")
        # print(f"T5处理 - batch_size: {batch_size}")
        # print(f"T5处理 - num_images_per_prompt: {num_images_per_prompt}")
        # print(f"T5处理 - max_sequence_length: {max_sequence_length}")
        # print(f"T5处理 - device: {device}")
        # print(f"T5处理 - dtype: {dtype}")

        if self.text_encoder_3 is None:
            #print("T5处理 - text_encoder_3为None，返回零张量")
            return torch.zeros(
                (
                    batch_size * num_images_per_prompt,
                    self.tokenizer_max_length,
                    self.transformer.config.joint_attention_dim,
                ),
                device=device,
                dtype=dtype,
            )

        text_inputs = self.tokenizer_3(
            prompt,
            padding="max_length",
            max_length=max_sequence_length,
            truncation=True,
           # add_special_tokens=True,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
        # if torch.isnan(text_input_ids).any():
        #     print("T5处理 - text_input_ids输入提示包含NaN值")
        # else:
        #     print("T5处理 - text_input_ids",text_input_ids)
        # print(f"T5处理 - text_input_ids形状: {text_input_ids.shape}")
        # print(f"T5处理 - text_input_ids范围: [{text_input_ids.min().item()}, {text_input_ids.max().item()}]")
        # print(f"T5处理 - tokenizer_3.vocab_size: {self.tokenizer_3.vocab_size}")
        
        # # 检查输入token IDs是否包含非法值
        # if torch.any(text_input_ids < 0):
        #     print(f"警告：发现负数token ID，最小值: {text_input_ids.min().item()}")
        # if torch.any(text_input_ids >= self.tokenizer_3.vocab_size):
        #     print(f"警告：发现超过词汇表大小的token ID，最大值: {text_input_ids.max().item()}, vocab_size: {self.tokenizer_3.vocab_size}")
        
        untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids

        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
            removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
            logger.warning(
                "The following part of your input was truncated because `max_sequence_length` is set to "
                f" {max_sequence_length} tokens: {removed_text}"
            )

        # 将输入移动到设备并确保数据类型正确
        text_input_ids = text_input_ids.to(device)#, dtype=torch.long)
        # print(f"T5处理 - text_input_ids形状: ",text_input_ids)
        # print(f"T5处理 - text_input_ids设备: {text_input_ids.device}, dtype: {text_input_ids.dtype}")
        
        # 检查text_encoder_3的状态
        # print(f"T5处理 - text_encoder_3设备: {next(self.text_encoder_3.parameters()).device}")
        # print(f"T5处理 - text_encoder_3.dtype: {self.text_encoder_3.dtype}")
        with torch.autocast(device.type if isinstance(device, torch.device) else "cuda", enabled=False):
            prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0]        
        #prompt_embeds = self.text_encoder_3(text_input_ids)[0]
        # print(f"T5处理 - T5编码器输出形状: {prompt_embeds.shape}")
        # print(f"T5处理 - T5编码器输出设备: {prompt_embeds.device}")
        # print(f"T5处理 - T5编码器输出dtype: {prompt_embeds.dtype}")
        
        # # 检查T5编码器输出是否包含NaN或inf
        # has_nan = torch.isnan(prompt_embeds).any()
        # has_inf = torch.isinf(prompt_embeds).any()
        # if has_nan or has_inf:
        #     print(f"警告：T5编码器输出包含NaN: {has_nan} 或inf: {has_inf}")
        #     print(f"T5编码器输出统计 - min: {prompt_embeds.min().item()}, max: {prompt_embeds.max().item()}, mean: {prompt_embeds.mean().item()}")
        #     if has_nan:
        #         nan_locations = torch.where(torch.isnan(prompt_embeds))
        #         print(f"NaN位置 - 前10个: {[(nan_locations[i][:10].tolist()) for i in range(len(nan_locations))]}")
        
        dtype = self.text_encoder_3.dtype
        # 强制在无 autocast 的上下文中运行 T5 以避免 fp16 溢出为 NaN

        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
        #print(f"T5处理 - 转换后prompt_embeds dtype: {prompt_embeds.dtype}, device: {prompt_embeds.device}")

        _, seq_len, _ = prompt_embeds.shape

        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
        # print(f"T5处理 - 最终输出形状: {prompt_embeds.shape}")
        
        # # 检查最终输出是否包含NaN
        # if torch.isnan(prompt_embeds).any():
        #     print(f"警告：T5最终输出包含NaN，位置: {torch.where(torch.isnan(prompt_embeds))}")
        #     print(f"T5最终输出统计 - min: {prompt_embeds.min().item()}, max: {prompt_embeds.max().item()}, mean: {prompt_embeds.mean().item()}")
        # print("最终输出",prompt_embeds)
        return prompt_embeds

    def _get_clip_prompt_embeds(
        self,
        prompt: Union[str, List[str]],
        num_images_per_prompt: int = 1,
        device: Optional[torch.device] = None,
        clip_skip: Optional[int] = None,
        clip_model_index: int = 0,
    ):
        device = device or self._execution_device

        clip_tokenizers = [self.tokenizer, self.tokenizer_2]
        clip_text_encoders = [self.text_encoder, self.text_encoder_2]

        tokenizer = clip_tokenizers[clip_model_index]
        text_encoder = clip_text_encoders[clip_model_index]
        
        # print(f"CLIP处理 - clip_model_index: {clip_model_index}")
        # print(f"CLIP处理 - 使用的tokenizer: {type(tokenizer)}")
        # print(f"CLIP处理 - 使用的text_encoder: {type(text_encoder)}")

        prompt = [prompt] if isinstance(prompt, str) else prompt
        batch_size = len(prompt)
        # print(f"CLIP处理 - 输入提示: {prompt}")
        # print(f"CLIP处理 - batch_size: {batch_size}")

        text_inputs = tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer_max_length,
            truncation=True,
            return_tensors="pt",
        )

        text_input_ids = text_inputs.input_ids
        # print(f"CLIP处理 - text_input_ids形状: {text_input_ids.shape}")
        # print(f"CLIP处理 - text_input_ids范围: [{text_input_ids.min().item()}, {text_input_ids.max().item()}]")
        
        untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
            removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
            logger.warning(
                "The following part of your input was truncated because CLIP can only handle sequences up to"
                f" {self.tokenizer_max_length} tokens: {removed_text}"
            )
        prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
        pooled_prompt_embeds = prompt_embeds[0]
        
        # print(f"CLIP处理 - pooled_prompt_embeds形状: {pooled_prompt_embeds.shape}")
        # print(f"CLIP处理 - pooled_prompt_embeds设备: {pooled_prompt_embeds.device}")
        # print(f"CLIP处理 - pooled_prompt_embeds dtype: {pooled_prompt_embeds.dtype}")
        
        # 检查CLIP编码器输出是否包含NaN或inf
        # has_nan = torch.isnan(pooled_prompt_embeds).any()
        # has_inf = torch.isinf(pooled_prompt_embeds).any()
        # if has_nan or has_inf:
        #     print(f"警告：CLIP编码器pooled输出包含NaN: {has_nan} 或inf: {has_inf}")
        #     print(f"CLIP编码器pooled输出统计 - min: {pooled_prompt_embeds.min().item()}, max: {pooled_prompt_embeds.max().item()}, mean: {pooled_prompt_embeds.mean().item()}")
        
        if clip_skip is None:
            prompt_embeds = prompt_embeds.hidden_states[-2]
        else:
            prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
        
        # 检查CLIP编码器embeds输出是否包含NaN或inf
        # has_nan = torch.isnan(prompt_embeds).any()
        # has_inf = torch.isinf(prompt_embeds).any()
        # if has_nan or has_inf:
        #     print(f"警告：CLIP编码器embeds输出包含NaN: {has_nan} 或inf: {has_inf}")
        #     print(f"CLIP编码器embeds输出统计 - min: {prompt_embeds.min().item()}, max: {prompt_embeds.max().item()}, mean: {prompt_embeds.mean().item()}")
        
        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
        #print(f"CLIP处理 - 转换后prompt_embeds dtype: {prompt_embeds.dtype}, device: {prompt_embeds.device}")

        _, seq_len, _ = prompt_embeds.shape
        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
        # print(f"CLIP处理 - 最终prompt_embeds形状: {prompt_embeds.shape}")
        
        # # 检查CLIP编码器embeds最终输出是否包含NaN
        # if torch.isnan(prompt_embeds).any():
        #     print(f"警告：CLIP编码器embeds最终输出包含NaN，位置: {torch.where(torch.isnan(prompt_embeds))}")
        #     print(f"CLIP编码器embeds最终输出统计 - min: {prompt_embeds.min().item()}, max: {prompt_embeds.max().item()}, mean: {prompt_embeds.mean().item()}")

        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1)
        pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1)
        # print(f"CLIP处理 - 最终pooled_prompt_embeds形状: {pooled_prompt_embeds.shape}")
        
        # # 检查CLIP编码器pooled最终输出是否包含NaN
        # if torch.isnan(pooled_prompt_embeds).any():
        #     print(f"警告：CLIP编码器pooled最终输出包含NaN，位置: {torch.where(torch.isnan(pooled_prompt_embeds))}")
        #     print(f"CLIP编码器pooled最终输出统计 - min: {pooled_prompt_embeds.min().item()}, max: {pooled_prompt_embeds.max().item()}, mean: {pooled_prompt_embeds.mean().item()}")

        return prompt_embeds, pooled_prompt_embeds

    def encode_prompt(
        self,
        prompt: Union[str, List[str]],
        prompt_2: Union[str, List[str]],
        prompt_3: Union[str, List[str]],
        device: Optional[torch.device] = None,
        num_images_per_prompt: int = 1,
        do_classifier_free_guidance: bool = True,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        negative_prompt_3: Optional[Union[str, List[str]]] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        clip_skip: Optional[int] = None,
        max_sequence_length: int = 256,
        lora_scale: Optional[float] = None,
    ):
        r"""

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in all text-encoders
            prompt_3 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
                used in all text-encoders
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`):
                whether to use classifier free guidance or not
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
            negative_prompt_3 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
                `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders.
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
            lora_scale (`float`, *optional*):
                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
        """
        device = device or self._execution_device
        # print(f"encode_prompt - 开始处理提示编码")
        # print(f"encode_prompt - device: {device}")
        # print(f"encode_prompt - num_images_per_prompt: {num_images_per_prompt}")
        # print(f"encode_prompt - do_classifier_free_guidance: {do_classifier_free_guidance}")
        # print(f"encode_prompt - max_sequence_length: {max_sequence_length}")

        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin):
            self._lora_scale = lora_scale

            # dynamically adjust the LoRA scale
            if self.text_encoder is not None and USE_PEFT_BACKEND:
                scale_lora_layers(self.text_encoder, lora_scale)
            if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
                scale_lora_layers(self.text_encoder_2, lora_scale)

        prompt = [prompt] if isinstance(prompt, str) else prompt
        if prompt is not None:
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]
       # print(f"encode_prompt - batch_size: {batch_size}")

        if prompt_embeds is None:
            prompt_2 = prompt_2 or prompt
            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2

            prompt_3 = prompt_3 or prompt
            prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3

            prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds(
                prompt=prompt,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
                clip_skip=clip_skip,
                clip_model_index=0,
            )
            prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds(
                prompt=prompt_2,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
                clip_skip=clip_skip,
                clip_model_index=1,
            )
            clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1)

            t5_prompt_embed = self._get_t5_prompt_embeds(
                prompt=prompt_3,
                num_images_per_prompt=num_images_per_prompt,
                max_sequence_length=max_sequence_length,
                device=device,
            )

            clip_prompt_embeds = torch.nn.functional.pad(
                clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1])
            )
            prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2)
            pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1)

        if do_classifier_free_guidance and negative_prompt_embeds is None:
            negative_prompt = negative_prompt or ""
            negative_prompt_2 = negative_prompt_2 or negative_prompt
            negative_prompt_3 = negative_prompt_3 or negative_prompt

            # normalize str to list
            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
            negative_prompt_2 = (
                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
            )
            negative_prompt_3 = (
                batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3
            )

            if prompt is not None and type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )

            negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds(
                negative_prompt,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
                clip_skip=None,
                clip_model_index=0,
            )
            negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds(
                negative_prompt_2,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
                clip_skip=None,
                clip_model_index=1,
            )
            negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1)

            t5_negative_prompt_embed = self._get_t5_prompt_embeds(
                prompt=negative_prompt_3,
                num_images_per_prompt=num_images_per_prompt,
                max_sequence_length=max_sequence_length,
                device=device,
            )

            negative_clip_prompt_embeds = torch.nn.functional.pad(
                negative_clip_prompt_embeds,
                (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]),
            )

            negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2)
            negative_pooled_prompt_embeds = torch.cat(
                [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1
            )

        if self.text_encoder is not None:
            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder, lora_scale)

        if self.text_encoder_2 is not None:
            if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder_2, lora_scale)

        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds

    def check_inputs(
        self,
        prompt,
        prompt_2,
        prompt_3,
        height,
        width,
        negative_prompt=None,
        negative_prompt_2=None,
        negative_prompt_3=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        pooled_prompt_embeds=None,
        negative_pooled_prompt_embeds=None,
        callback_on_step_end_tensor_inputs=None,
        max_sequence_length=None,
    ):
        if (
            height % (self.vae_scale_factor * self.patch_size) != 0
            or width % (self.vae_scale_factor * self.patch_size) != 0
        ):
            raise ValueError(
                f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}."
                f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}."
            )

        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt_2 is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt_3 is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
        elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)):
            raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}")

        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )
        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )
        elif negative_prompt_3 is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

        if prompt_embeds is not None and pooled_prompt_embeds is None:
            raise ValueError(
                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
            )

        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
            raise ValueError(
                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
            )

        if max_sequence_length is not None and max_sequence_length > 512:
            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")

    def prepare_latents(
        self,
        batch_size,
        num_channels_latents,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
    ):
        if latents is not None:
            return latents.to(device=device, dtype=dtype)

        shape = (
            batch_size,
            num_channels_latents,
            int(height) // self.vae_scale_factor,
            int(width) // self.vae_scale_factor,
        )

        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)

        return latents

    @property
    def guidance_scale(self):
        return self._guidance_scale

    @property
    def skip_guidance_layers(self):
        return self._skip_guidance_layers

    @property
    def clip_skip(self):
        return self._clip_skip

    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
    # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
    # corresponds to doing no classifier free guidance.
    @property
    def do_classifier_free_guidance(self):
        return self._guidance_scale > 1

    @property
    def joint_attention_kwargs(self):
        return self._joint_attention_kwargs

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @property
    def interrupt(self):
        return self._interrupt

    # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_image
    def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor:
        """Encodes the given image into a feature representation using a pre-trained image encoder.

        Args:
            image (`PipelineImageInput`):
                Input image to be encoded.
            device: (`torch.device`):
                Torch device.

        Returns:
            `torch.Tensor`: The encoded image feature representation.
        """
        if not isinstance(image, torch.Tensor):
            image = self.feature_extractor(image, return_tensors="pt").pixel_values

        image = image.to(device=device, dtype=self.dtype)

        return self.image_encoder(image, output_hidden_states=True).hidden_states[-2]

    # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.prepare_ip_adapter_image_embeds
    def prepare_ip_adapter_image_embeds(
        self,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        ip_adapter_image_embeds: Optional[torch.Tensor] = None,
        device: Optional[torch.device] = None,
        num_images_per_prompt: int = 1,
        do_classifier_free_guidance: bool = True,
    ) -> torch.Tensor:
        """Prepares image embeddings for use in the IP-Adapter.

        Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed.

        Args:
            ip_adapter_image (`PipelineImageInput`, *optional*):
                The input image to extract features from for IP-Adapter.
            ip_adapter_image_embeds (`torch.Tensor`, *optional*):
                Precomputed image embeddings.
            device: (`torch.device`, *optional*):
                Torch device.
            num_images_per_prompt (`int`, defaults to 1):
                Number of images that should be generated per prompt.
            do_classifier_free_guidance (`bool`, defaults to True):
                Whether to use classifier free guidance or not.
        """
        device = device or self._execution_device

        if ip_adapter_image_embeds is not None:
            if do_classifier_free_guidance:
                single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2)
            else:
                single_image_embeds = ip_adapter_image_embeds
        elif ip_adapter_image is not None:
            single_image_embeds = self.encode_image(ip_adapter_image, device)
            if do_classifier_free_guidance:
                single_negative_image_embeds = torch.zeros_like(single_image_embeds)
        else:
            raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.")

        image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)

        if do_classifier_free_guidance:
            negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0)
            image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0)

        return image_embeds.to(device=device)

    def enable_sequential_cpu_offload(self, *args, **kwargs):
        if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload:
            logger.warning(
                "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses "
                "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling "
                "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`."
            )

        super().enable_sequential_cpu_offload(*args, **kwargs)

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        prompt_3: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 28,
        sigmas: Optional[List[float]] = None,
        guidance_scale: float = 7.0,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        negative_prompt_3: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        ip_adapter_image_embeds: Optional[torch.Tensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
        clip_skip: Optional[int] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        max_sequence_length: int = 256,
        skip_guidance_layers: List[int] = None,
        skip_layer_guidance_scale: float = 2.8,
        skip_layer_guidance_stop: float = 0.2,
        skip_layer_guidance_start: float = 0.01,
        mu: Optional[float] = None,
        model: Optional[Any] = None,  # 添加 model 参数，默认为 None
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                will be used instead
            prompt_3 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
                will be used instead
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image. This is set to 1024 by default for the best results.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image. This is set to 1024 by default for the best results.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            sigmas (`List[float]`, *optional*):
                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                will be used.
            guidance_scale (`float`, *optional*, defaults to 7.0):
                Guidance scale as defined in [Classifier-Free Diffusion
                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
                of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
                `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
                the text `prompt`, usually at the expense of lower image quality.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used instead
            negative_prompt_3 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
                `text_encoder_3`. If not defined, `negative_prompt` is used instead
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            ip_adapter_image (`PipelineImageInput`, *optional*):
                Optional image input to work with IP Adapters.
            ip_adapter_image_embeds (`torch.Tensor`, *optional*):
                Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images,
                emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to
                `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of
                a plain tuple.
            joint_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.
            max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`.
            skip_guidance_layers (`List[int]`, *optional*):
                A list of integers that specify layers to skip during guidance. If not provided, all layers will be
                used for guidance. If provided, the guidance will only be applied to the layers specified in the list.
                Recommended value by StabiltyAI for Stable Diffusion 3.5 Medium is [7, 8, 9].
            skip_layer_guidance_scale (`int`, *optional*): The scale of the guidance for the layers specified in
                `skip_guidance_layers`. The guidance will be applied to the layers specified in `skip_guidance_layers`
                with a scale of `skip_layer_guidance_scale`. The guidance will be applied to the rest of the layers
                with a scale of `1`.
            skip_layer_guidance_stop (`int`, *optional*): The step at which the guidance for the layers specified in
                `skip_guidance_layers` will stop. The guidance will be applied to the layers specified in
                `skip_guidance_layers` until the fraction specified in `skip_layer_guidance_stop`. Recommended value by
                StabiltyAI for Stable Diffusion 3.5 Medium is 0.2.
            skip_layer_guidance_start (`int`, *optional*): The step at which the guidance for the layers specified in
                `skip_guidance_layers` will start. The guidance will be applied to the layers specified in
                `skip_guidance_layers` from the fraction specified in `skip_layer_guidance_start`. Recommended value by
                StabiltyAI for Stable Diffusion 3.5 Medium is 0.01.
            mu (`float`, *optional*): `mu` value used for `dynamic_shifting`.
            model (`SD3WithRectifiedNoise`, *optional*): 
                Optional SD3WithRectifiedNoise model for enhanced noise prediction. If provided, will be used instead of 
                the default transformer for denoising. The model should be an instance of SD3WithRectifiedNoise class.

        Examples:

        Returns:
            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a
            `tuple`. When returning a tuple, the first element is a list with the generated images.
        """

        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor
        #height=512
        #width=512
        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            prompt_2,
            prompt_3,
            height,
            width,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            negative_prompt_3=negative_prompt_3,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
            max_sequence_length=max_sequence_length,
        )

        self._guidance_scale = guidance_scale
        self._skip_layer_guidance_scale = skip_layer_guidance_scale
        self._clip_skip = clip_skip
        self._joint_attention_kwargs = joint_attention_kwargs
        self._interrupt = False

        # 2. Define call parameters
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        device = self._execution_device

        lora_scale = (
            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
        )
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            prompt_3=prompt_3,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            negative_prompt_3=negative_prompt_3,
            do_classifier_free_guidance=self.do_classifier_free_guidance,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            device=device,
            clip_skip=self.clip_skip,
            num_images_per_prompt=num_images_per_prompt,
            max_sequence_length=max_sequence_length,
            lora_scale=lora_scale,
        )

        if self.do_classifier_free_guidance:
            if skip_guidance_layers is not None:
                original_prompt_embeds = prompt_embeds
                original_pooled_prompt_embeds = pooled_prompt_embeds
           # print("检测negative_prompt_embeds",negative_prompt_embeds)
            #print("检测pooled_prompt_embeds",prompt_embeds)
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)

            pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)

        # 4. Prepare latent variables
        num_channels_latents = self.transformer.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 5. Prepare timesteps
        scheduler_kwargs = {}
        if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None:
            _, _, height, width = latents.shape
            image_seq_len = (height // self.transformer.config.patch_size) * (
                width // self.transformer.config.patch_size
            )
            mu = calculate_shift(
                image_seq_len,
                self.scheduler.config.get("base_image_seq_len", 256),
                self.scheduler.config.get("max_image_seq_len", 4096),
                self.scheduler.config.get("base_shift", 0.5),
                self.scheduler.config.get("max_shift", 1.16),
            )
            scheduler_kwargs["mu"] = mu
        elif mu is not None:
            scheduler_kwargs["mu"] = mu
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            sigmas=sigmas,
            **scheduler_kwargs,
        )
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
        self._num_timesteps = len(timesteps)

        # 6. Prepare image embeddings
        if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None:
            ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds(
                ip_adapter_image,
                ip_adapter_image_embeds,
                device,
                batch_size * num_images_per_prompt,
                self.do_classifier_free_guidance,
            )

            if self.joint_attention_kwargs is None:
                self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds}
            else:
                self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds)

        # 7. Denoising loop
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                # expand the latents if we are doing classifier free guidance
                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand(latent_model_input.shape[0])

                # Check for NaN in latents before transformer
                if torch.isnan(latents).any():
                    # print(f"NaN detected in latents at step {i}")
                    # print(f"NaN locations: {torch.where(torch.isnan(latents))}")
                    break
                # 优先使用传入的 model 参数，其次使用类属性中的 model，最后回退到默认 transformer
                effective_model = model or getattr(self, 'model', None) or self.transformer
                
                if hasattr(effective_model, '__call__') and callable(effective_model):
                    # 使用有效的模型进行预测
                    # 检查模型是否支持 skip_layers 参数
                    if hasattr(effective_model, 'forward') and 'skip_layers' in inspect.signature(effective_model.forward).parameters:
                        noise_pred_output = effective_model(
                            hidden_states=latent_model_input,
                            timestep=timestep,
                            encoder_hidden_states=prompt_embeds,
                            pooled_projections=pooled_prompt_embeds,
                            #joint_attention_kwargs=self.joint_attention_kwargs,
                            return_dict=False,
                            #skip_layers=skip_guidance_layers if skip_guidance_layers is not None else None,
                        )
                        #print(f"effective_model type: {type(effective_model)}")
                        #print(f"noise_pred_output: {noise_pred_output}")
                    else:
                        # SD3WithRectifiedNoise 不支持 skip_layers 参数
                        noise_pred_output = effective_model(
                            hidden_states=latent_model_input,
                            timestep=timestep,
                            encoder_hidden_states=prompt_embeds,
                            pooled_projections=pooled_prompt_embeds,
                            joint_attention_kwargs=self.joint_attention_kwargs,
                            return_dict=False,
                        )
                    # 正确处理 SD3WithRectifiedNoise 模型的输出
                    # SD3WithRectifiedNoise 返回 (final_output, mean_out, var_out) 元组，我们只需要第一个元素
                    # 如果返回的是字典，则使用 "sample" 键
                    if isinstance(noise_pred_output, dict):
                        noise_pred = noise_pred_output["sample"]
                    elif isinstance(noise_pred_output, tuple):
                        # 对于 SD3WithRectifiedNoise，取第一个输出作为主要预测结果
                        noise_pred = noise_pred_output[0]
                    else:
                        noise_pred = noise_pred_output
                else:
                    # 使用默认的 transformer 进行预测
                    noise_pred = self.transformer(
                        hidden_states=latent_model_input,
                        timestep=timestep,
                        encoder_hidden_states=prompt_embeds,
                        pooled_projections=pooled_prompt_embeds,
                        joint_attention_kwargs=self.joint_attention_kwargs,
                        return_dict=False,
                    )[0]

                # Check for NaN in noise prediction
                if torch.isnan(noise_pred).any():
                    # print(f"NaN detected in noise_pred at step {i}")
                    # print(f"NaN locations: {torch.where(torch.isnan(noise_pred))}")
                    # print(f"noise_pred stats - min: {noise_pred.min().item()}, max: {noise_pred.max().item()}, mean: {noise_pred.mean().item()}")
                    break

                # perform guidance
                if self.do_classifier_free_guidance:
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
                    
                    # Check for NaN after guidance
                    if torch.isnan(noise_pred).any():
                        # print(f"NaN detected in noise_pred after guidance at step {i}")
                        # print(f"noise_pred_uncond stats - min: {noise_pred_uncond.min().item()}, max: {noise_pred_uncond.max().item()}, mean: {noise_pred_uncond.mean().item()}")
                        # print(f"noise_pred_text stats - min: {noise_pred_text.min().item()}, max: {noise_pred_text.max().item()}, mean: {noise_pred_text.mean().item()}")
                        # print(f"guidance_scale: {self.guidance_scale}")
                        break
                    
                    should_skip_layers = (
                        True
                        if i > num_inference_steps * skip_layer_guidance_start
                        and i < num_inference_steps * skip_layer_guidance_stop
                        else False
                    )
                    if skip_guidance_layers is not None and should_skip_layers:
                        timestep = t.expand(latents.shape[0])
                        latent_model_input = latents
                        # 修改 skip_guidance_layers 部分的 transformer 调用逻辑以支持 SD3WithRectifiedNoise 模型
                        # 优先使用传入的 model 参数，其次使用类属性中的 model，最后回退到默认 transformer
                        effective_model = model or getattr(self, 'model', None) or self.transformer
                            
                        if hasattr(effective_model, '__call__') and callable(effective_model):
                            # 使用有效的模型进行预测
                            # 检查模型是否支持 skip_layers 参数
                            if hasattr(effective_model, 'forward') and 'skip_layers' in inspect.signature(effective_model.forward).parameters:
                                noise_pred_skip_output = effective_model(
                                    hidden_states=latent_model_input,
                                    timestep=timestep,
                                    encoder_hidden_states=original_prompt_embeds,
                                    pooled_projections=original_pooled_prompt_embeds,
                                    joint_attention_kwargs=self.joint_attention_kwargs,
                                    return_dict=False,
                                    skip_layers=skip_guidance_layers,
                                )
                            else:
                                # SD3WithRectifiedNoise 不支持 skip_layers 参数
                                noise_pred_skip_output = effective_model(
                                    hidden_states=latent_model_input,
                                    timestep=timestep,
                                    encoder_hidden_states=original_prompt_embeds,
                                    pooled_projections=original_pooled_prompt_embeds,
                                    joint_attention_kwargs=self.joint_attention_kwargs,
                                    return_dict=False,
                                )
                            # 正确处理 SD3WithRectifiedNoise 模型的输出
                            # SD3WithRectifiedNoise 返回 (final_output, mean_out, var_out) 元组，我们只需要第一个元素
                            # 如果返回的是字典，则使用 "sample" 键
                            if isinstance(noise_pred_skip_output, dict):
                                noise_pred_skip_layers = noise_pred_skip_output["sample"]
                            elif isinstance(noise_pred_skip_output, tuple):
                                # 对于 SD3WithRectifiedNoise，取第一个输出作为主要预测结果
                                noise_pred_skip_layers = noise_pred_skip_output[0]
                            else:
                                noise_pred_skip_layers = noise_pred_skip_output
                        else:
                            # 使用默认的 transformer 进行预测
                            noise_pred_skip_layers = self.transformer(
                                hidden_states=latent_model_input,
                                timestep=timestep,
                                encoder_hidden_states=original_prompt_embeds,
                                pooled_projections=original_pooled_prompt_embeds,
                                joint_attention_kwargs=self.joint_attention_kwargs,
                                return_dict=False,
                                skip_layers=skip_guidance_layers,
                            )[0]
                        
                        # Check for NaN in skip layers noise prediction
                        if torch.isnan(noise_pred_skip_layers).any():
                            # print(f"NaN detected in noise_pred_skip_layers at step {i}")
                            # print(f"NaN locations: {torch.where(torch.isnan(noise_pred_skip_layers))}")
                            break
                        
                        noise_pred = (
                            noise_pred + (noise_pred_text - noise_pred_skip_layers) * self._skip_layer_guidance_scale
                        )
                        
                        # Check for NaN after skip layer guidance
                        if torch.isnan(noise_pred).any():
                            # print(f"NaN detected in noise_pred after skip layer guidance at step {i}")
                            break

                # compute the previous noisy sample x_t -> x_t-1
                latents_dtype = latents.dtype
                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]

                # Check for NaN in latents after scheduler step
                if torch.isnan(latents).any():
                    # print(f"NaN detected in latents after scheduler step at step {i}")
                    # print(f"noise_pred stats - min: {noise_pred.min().item()}, max: {noise_pred.max().item()}, mean: {noise_pred.mean().item()}")
                    break

                # Print intermediate results
                # print(f"Step {i+1}/{num_inference_steps}, Timestep: {t.item():.2f}, Latents mean: {latents.mean().item():.6f}, Latents std: {latents.std().item():.6f}")

                if latents.dtype != latents_dtype:
                    if torch.backends.mps.is_available():
                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
                        latents = latents.to(latents_dtype)

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                    pooled_prompt_embeds = callback_outputs.pop("pooled_prompt_embeds", pooled_prompt_embeds)

                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()

                if XLA_AVAILABLE:
                    xm.mark_step()

        if output_type == "latent":
            image = latents

        else:
            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor

            image = self.vae.decode(latents, return_dict=False)[0]
            image = self.image_processor.postprocess(image, output_type=output_type)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return StableDiffusion3PipelineOutput(images=image)